Resumo:
Aeroderivative gas turbines are widely used on Floating Production Storage and Offloading (FPSO) oil and gas production platforms. These units serve as production, storage, and oil transfer facilities. These turbines are employed either for generating electrical power when connected to an electrical generator or for mechanically driving pumps and compressors when connected to these machines. Oil and gas production is intermittent and demands gas turbines to be available and flexible for safe and reliable operation. To achieve such conditions, monitoring them is a crucial factor to ensure operational safety. Another factor is diagnosing turbine faults, efficiently and reliably identifying where these faults are occurring, so that maintenance planning is effective in keeping the machine operational. In this context, the primary objective of this work is to develop a method to assist in diagnosing faults that affect the performance of aeroderivative gas turbines. To achieve this, a model of an aeroderivative gas turbine, consisting of a two-spool gas generator and a power turbine (PT), was developed using MATLAB/Simulink with the assistance of the T-MATS library. The developed model can represent the behaviour of the gas turbine when operating in a steady-state since the goal is to generate data for different operating conditions and different environmental conditions for subsequent use in a neural network model. The gas turbine model exhibited satisfactory behaviour, with deviations not exceeding 1% when compared to actual operational data. Subsequently, fault conditions were imposed on the model, which provided information about operational parameters of the turbogenerator operating with degradation in some of its components. The data generated by the gas turbine model in MATLAB/Simulink were used to feed a machine learning model for fault diagnosis. The proposed model consists of two parallel feedforward neural networks, one for regression and one for classification. The regression network aims to handle numerical values and the turbine's behaviour, while the classification network aims to identify and classify faults. Both networks performed well for both single and combined fault problems, with mean squared errors not exceeding 5x10-6 for the regression network and a percentage error of 0.37% for the classification network.
Key
